GPT-56T 861 —
MUSE-SPK 837 —
GPT-56SC 789 -0.1%
GLM-5 781 —
CL-OP55X 779 -0.1%
GROK-46H 779 -0.1%
QWEN-38X 748 —
GPT-6A 743 —
KIMI-K3X 742 —
CL-FAB5H 697 -0.1%
CL-OP5H 674 -0.1%
GEM-38FH 672 —
CL-OP5X 669 -0.1%
CL-OP55H 667 -0.1%
CL-OP46H 656 -0.2%
CL-OP47H 647 -0.2%
GPT-56S 617 -0.2%
GEM-37FH 609 -0.2%
GEM-36FH 592 -0.2%
CL-OP48H 587 -0.2%
CL-OP47 580 -0.2%
GEM-35FH 579 -0.2%
GPT-55H 540 -0.2%
INKL 531 —
GEM-31P 511 -0.2%
CL-OP46 498 —
GEM-3P 498 —
CL-OP48 492 —
GPT-52 464 —
GPT-55 423 —
GPT-56T 861 —
MUSE-SPK 837 —
GPT-56SC 789 -0.1%
GLM-5 781 —
CL-OP55X 779 -0.1%
GROK-46H 779 -0.1%
QWEN-38X 748 —
GPT-6A 743 —
KIMI-K3X 742 —
CL-FAB5H 697 -0.1%
CL-OP5H 674 -0.1%
GEM-38FH 672 —
CL-OP5X 669 -0.1%
CL-OP55H 667 -0.1%
CL-OP46H 656 -0.2%
CL-OP47H 647 -0.2%
GPT-56S 617 -0.2%
GEM-37FH 609 -0.2%
GEM-36FH 592 -0.2%
CL-OP48H 587 -0.2%
CL-OP47 580 -0.2%
GEM-35FH 579 -0.2%
GPT-55H 540 -0.2%
INKL 531 —
GEM-31P 511 -0.2%
CL-OP46 498 —
GEM-3P 498 —
CL-OP48 492 —
GPT-52 464 —
GPT-55 423 —
← Back to feed

GPT-5.4 and Claude Opus 4.6 Tie at 81.8% on Terminal-Bench 2.0 — Scaffold Engineering Now Determines the Winner

The Terminal-Bench 2.0 leaderboard — 123 entries across real-world terminal task resolution — has converged on a result that reframes the model selection debate: with the same scaffolding, the top two models are statistically indistinguishable.

The Numbers

RankAgentModelOrgScore
1ForgeCodeGPT-5.4OpenAI81.8% ±2.0
2ForgeCodeClaude Opus 4.6Anthropic81.8% ±1.7
3TongAgentsGemini 3.1 ProGoogle (BIGAI)80.2% ±2.6
4SageAgentGPT-5.3-CodexOpenSage78.4% ±2.2
5ForgeCodeGemini 3.1 ProGoogle78.4% ±1.8
6DroidGPT-5.3-CodexFactory77.3% ±2.2
7CapyClaude Opus 4.6Capy75.3% ±2.4
8Simple CodexGPT-5.3-CodexOpenAI75.1% ±2.4

Positions 1 and 2 are the same scaffold wrapping two different models. The 6.5-point gap between ForgeCode-wrapped Claude Opus 4.6 (81.8%) and Capy-wrapped Claude Opus 4.6 (75.3%) is larger than any model-to-model gap in the top 5. The scaffold is doing the work.

What ForgeCode Is Doing Differently

ForgeCode is a purpose-built agent scaffolding optimised for Terminal-Bench’s task distribution — shell navigation, file manipulation, environment setup, multi-step command sequencing. Its architecture prioritises tight error recovery loops and explicit state tracking between tool calls.

The contrast with simpler scaffolds like Capy (75.3% with the same Opus 4.6) or OpenAI’s own Simple Codex baseline (75.1% with GPT-5.3-Codex) shows that 6–7 percentage points of measured capability are an engineering problem, not a model problem.

Gemini’s Competitive Position

Gemini 3.1 Pro at 80.2% via TongAgents (BIGAI) lands 1.6 points behind the top tier — within two standard deviations given the confidence intervals. Under ForgeCode directly, Gemini 3.1 Pro reaches 78.4%, suggesting TongAgents’ architecture suits Gemini’s tool-use patterns better than a generic scaffold does.

The implication: a lab that invests in scaffold engineering optimised for a specific model can close most of the gap to the frontier. The model matters, but it is no longer the whole story.

What This Means for Enterprise AI Adoption

Enterprises evaluating frontier models for agentic coding pipelines face a compound decision: which model to use, and which scaffold or orchestration framework to deploy it with. Terminal-Bench 2.0 suggests that choosing a weaker scaffold with the strongest model will under-perform a well-engineered scaffold on a second-tier model.

The practical takeaway: at current frontier capability levels, switching from GPT-5.4 to Claude Opus 4.6 (or vice versa) within the same scaffold produces no measurable outcome difference. The scaffold vendor is now a first-order variable in the build decision.